Enabling On-Device Learning with Deep Spiking Neural Networks for Speech Recognition

Spiking recurrent neural networks are gaining traction in solving complex temporal tasks. In general, spiking neural networks are resilient and computationally powerful. These intrinsic properties make them attractive for learning on edge devices. In this work, we propose a semi-supervised deep spiking neural network (deep-liquid state machine) that can be deployed on embedded devices. We demonstrate a high-level memristor based neuromorphic architecture for the proposed deep spiking network. An experimental TiN-TaOx-TaTiN memristor device stack model is used for analyzing the overall architecture performance. An accuracy of 95.83 +/- 1.74% is achieved for a speech recognition task, on the standard TIMIT dataset. To study the robustness of the proposed architecture, the accuracy of the network was tested with and without noise in the memristor devices.